CN111161159A - Image defogging method and device based on combination of priori knowledge and deep learning - Google Patents

Abstract

The invention relates to an image defogging method and device based on combination of priori knowledge and deep learning. The method comprises the following steps: decomposing the input original foggy image Z by using a weighted guide filter to obtain a detail layer image Z_eAnd base layer image Z_b(ii) a Using a quadtree search method to search the base layer image Z_bProcessing to obtain a global atmosphere light component image A_cWherein c belongs to { r, g, b }; constructing a deep convolutional neural network on the base layer image Z_bProcessing is carried out to obtain a transmissivity image t; utilizing the global atmospheric light component image A based on an atmospheric scattering model_cThe transmittance image t and the detail layer image Z_eAnd restoring the original foggy image Z to obtain a defogged image J.

Description

Image defogging method and device based on combination of priori knowledge and deep learning

Technical Field

The invention relates to the technical field of image processing, in particular to an image defogging method and device based on combination of priori knowledge and deep learning.

Background

In the foggy weather, floating particles in the sun, fog or dust cause the picture to fade and be blurred, the contrast and softness are reduced, the image quality is severely restricted, and the applications of video monitoring and analysis, target identification, urban traffic, aerial photography, military and national defense and the like are limited. Therefore, the method has a direct relationship with the civil life to the clear processing of the foggy image, and has great practical significance to the production and the life of people.

The existing defogging algorithms are mainly divided into three categories: non-model based defogging algorithms, and deep learning based defogging algorithms. The defogging algorithm based on the non-model mainly achieves the aim of improving the image by directly stretching the contrast of the image. The common methods are as follows: histogram equalization, homomorphic filtering algorithms, Retinex model-based algorithms, and Retinex model-improved-based algorithms. The methods carry out defogging according to the optical imaging principle, so that the contrast among image colors is more balanced, the image effect is softer, but the obtained image cannot be effectively enhanced in contrast, the method weakens dark or bright areas in the original image, and blurs the key points of the image.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides an image defogging method and device based on combination of priori knowledge and deep learning.

In a first aspect, the present invention provides an image defogging method based on a combination of priori knowledge and deep learning, including:

decomposing the input original foggy image Z by using a weighted guide filter to obtain a detail layer image Z_eAnd base layer image Z_b；

Using a quadtree search method to search the base layer image Z_bProcessing to obtain a global atmosphere light component image A_cWherein c belongs to { r, g, b };

constructing a deep convolutional neural network pair to the base layerImage Z_bProcessing is carried out to obtain a transmissivity image t;

utilizing the global atmospheric light component image A based on an atmospheric scattering model_cThe transmittance image t and the detail layer image Z_eAnd restoring the original foggy image Z to obtain a defogged image J.

Preferably, the input original fog image Z is decomposed by a weighted guided filter to obtain a base layer image Z_bThe method specifically comprises the following steps:

acquiring a pixel gray value Z (x, y) at any point (x, y) in the original foggy image Z;

obtaining a weighted filter coefficient a at the point (x, y) in the Z of the original foggy image by using the following formula_p(x, y) and b_p(x，y)：

Wherein, mu_Z,ρ(x, y) is the mean of the pixel grey values of the window centered at point (x, y) in the original hazy image Z, with ρ being the radius;

is the variance of the pixel gray value of the window with point (x, y) as the center and rho as the radius in the original foggy image Z; gamma-shaped_Y(x, Y) is the luminance component of the original hazy image Z at point (x, Y), Y ═ max { Z {_r,Z_g,Z_b}，Z_r,Z_g,Z_bR, G, B values at the point (x, y) in the hazy image Z, respectively; λ is a constant greater than 1;

obtaining the base layer image Z using the following formula_b：

Z_b(x,y)＝a_p(x,y)Z(x,y)+b_p(x,y)；

Wherein Z is_b(x, y) is the base layer image Z_bOf the pixel grey value at point (x, y).

Preferably, the base layer image Z is searched by using a quadtree search method_bProcessing to obtain global atmosphere lightPartial image A_cThe method specifically comprises the following steps:

step A, base layer image Z_bAre equally divided into four rectangular areas Z_b-i(i∈{1,2,3,4})，Z_b-iRespectively has a length and a width of Z_b1/2 for the length and width of;

b, defining the fraction of each rectangular area as the average value of the pixel gray value minus the standard deviation of the pixel gray value in the area;

step C, selecting the rectangular area with the highest score and further dividing the rectangular area into four rectangular areas;

step D, repeating the steps B and C, and carrying out iterative updating on the rectangular area with the highest score for n times to obtain a final subdivision area Z_b-end；

Obtaining the atmospheric light component image A by using the following formula_c:

|A_c(c∈{r,g,b})|＝min|((Z_b-end-r(x,y),Z_b-end-g(x,y),Z_b-end-b(x,y))-(255,255,255))|，

In the formula (Z)_b-end-r(x,y),Z_b-end-g(x,y),Z_b-end-b(x, y)) is the final subdivided region Z_b-endThe color vector at the midpoint (x, y), (255 ) is a pure white vector.

Preferably, the deep convolutional neural network is constructed on the base layer image Z_bThe transmittance image t is obtained by processing, and the method specifically comprises the following steps:

using the first convolution layer to pair the basic image Z_bDown-sampling to obtain low-resolution base layer image Z_b' post-extracting low-level features, wherein the base image Z is aligned using a first convolution layer_bThe formula for downsampling is:

wherein the content of the first and second substances,

convolution for the ith layer in the first convolution layerLayer at the index of channel c, the low resolution base layer image Z_bThe low-level feature of'; x is a base image Z_bY is the base image Z_bThe ordinate of (a); x' is a low resolution base layer image Z_b'abscissa, y' is the low resolution base layer image Z_bThe ordinate of `;

convolution weight arrays of the ith convolution layer in the first convolution layer under the index channels of the layers c and c';

a deviation vector of the ith convolution layer in the first convolution layer under the c layer index channel is obtained; σ (·) max (·,0) denotes a ReLU activation function and zero padding is used as a boundary condition for all of the first convolutional layers; s is the step length of the convolution kernel of the first convolution layer;

low layer characteristics to be obtained

The number of input layers is n_LExtracting a local feature L from a second convolution layer of 2, wherein the size of a convolution kernel in the second convolution layer is 3 multiplied by 3, and the step size is 1;

low layer feature obtained by laminating a second convolution layer

The number of input layers is n_G1After the third convolution layer with convolution kernel size of 3 × 3 and step length of 2 is input, the number of layers n is input_G2Obtaining a global feature G of the 2 full connection layer;

adding the local feature L and the global feature G, and inputting the sum into an activation function to obtain a low-level feature

Corresponding hybrid feature map F_L＝σ(L+G)；

For mixed feature map F_LObtaining sum-low by performing convolution processing for σ (L + G)Layer characteristics

And (3) performing up-sampling on the corresponding preliminary atmospheric refractive index characteristic graph, and outputting to obtain a transmittance image t.

Preferably, the global atmospheric light component image A is utilized based on an atmospheric scattering model_cThe transmittance image t and the detail layer image Z_eAnd recovering the original foggy image Z, wherein the formula for obtaining the defogged image J is as follows:

wherein J (x, y) is the gray value of the pixel at the point (x, y) in the defogged image J, and Z_e(x, y) is the detail layer image Z_eThe gray value of the pixel at the middle point (x, y), Z_b(x, y) is the base layer image Z_bA pixel gray value at a midpoint (x, y), t (x, y) being a transmittance at a midpoint (x, y) of the transmittance image t,

where η is a constant greater than zero.

Preferably, the pair of mixed feature maps F_LObtaining the feature of the lower layer by convolution processing with sigma (L + G)

And then, performing up-sampling on the atmospheric refractive index characteristic diagram, and outputting a transmittance image t, wherein the method further comprises the following steps:

the loss function is constructed as follows:

L＝L_r+w_cL_c

wherein L is_rTo reconstruct the loss function, L_cAs a function of color loss, w_cTo be assigned to the color loss function L_cWeight of, L_rIs shown as

L_cIs shown as

J is a defogged image, Z is a fogging image, c belongs to (R, G, B) as a channel index, and ∠ (J (x, y), Z (x, y)) represents the included angle of the three-dimensional color vectors of the fogging image and the defogged image at a pixel point (x, y);

and performing parameter adjustment processing on the deep convolutional neural network by using the loss function.

In a second aspect, the invention provides an image defogging device based on combination of priori knowledge and deep learning, wherein the device comprises a memory and a processor;

the memory for storing a computer program;

the processor, when executing the computer program, is configured to implement the image defogging method according to any one of claims 1 to 6 based on the combination of the prior knowledge and the deep learning.

The image defogging method and the image defogging device based on the combination of the prior knowledge and the deep learning have the advantages that the weighted guide filter is used for decomposing the foggy image to obtain a detail layer image and a basic layer image, then a quad-tree search method and a deep neural network are used for processing the basic layer image to obtain a global atmosphere light component image and a transmissivity image, and finally the global atmosphere light component image, the transmissivity image and the detail layer image are used for restoring the foggy image to obtain the defogged image. Because the global atmosphere light component image and the transmissivity image are estimated or calculated by utilizing the base layer image, the amplification of image noise is avoided, and the defogging treatment can be better carried out on the foggy image.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.

Fig. 1 is a schematic flow chart of an image defogging method based on combination of a priori knowledge and deep learning according to an embodiment of the present invention.

Detailed Description

The principles and features of this invention are described below in conjunction with the following drawings, which are set forth by way of illustration only and are not intended to limit the scope of the invention.

Referring to fig. 1, the single-image defogging method based on the combination of the prior knowledge and the deep learning described in the present invention includes the following steps:

decomposing the input original foggy image Z by using a weighted guide filter to obtain a detail layer image Z_eAnd base layer image Z_b；

Using a quadtree search method to search the base layer image Z_bProcessing to obtain a global atmosphere light component image A_c；

Constructing a deep convolutional neural network on the base layer image Z_bProcessing is carried out to obtain a transmissivity image t;

utilizing the global atmospheric light component image A based on an atmospheric scattering model_cThe transmittance image t and the detail layer image Z_eAnd restoring the original foggy image Z to obtain a defogged image J.

Specifically, the original input foggy image Z is decomposed by using a weighted guide filter to obtain a detail layer image Z_eAnd base layer image Z_bThe method specifically comprises the following steps:

acquiring a pixel gray value Z (x, y) at any point (x, y) in the original foggy image Z;

obtaining a weighted filter coefficient a at the point (x, y) in the Z of the original foggy image by using the following formula_p(x, y) and b_p(x，y)：

b_p(x，y)＝(1-a_p(x，y))μ_Z,ρ(x,y)；

Wherein, mu_Z,ρ(x, y) is the mean of the pixel grey values of the window centered at point (x, y) in the original hazy image Z, with ρ being the radius;

is the variance of the pixel gray value of the window with point (x, y) as the center and rho as the radius in the original foggy image Z;

Γ_Y(x, y) is the luminance component of point (x, y) of the original hazy image Z;

λ is a constant greater than 1;

obtaining the base layer image Z using the following formula_b：

Z_b(x,y)＝a_p(x,y)Z(x,y)+b_p(x,y)；

Wherein Z is_b(x, y) is the base layer image Z_bThe pixel gray value at point (x, y);

z (x, y) is a pixel gray value at any point (x, y) in the original foggy image Z;

a_p(x, y) and b_p(x, y) weighted filter coefficients at point (x, y) in the original hazy image Z.

Specifically, the base layer image Z is searched by using a quadtree search method_bProcessing to obtain a global atmosphere light component image A_cThe method specifically comprises the following steps:

step A, base layer image Z_bAre equally divided into four rectangular areas Z_b-i(i∈{1,2,3,4})，Z_b-iRespectively has a length and a width of Z_b1/2 for the length and width of;

b, defining the fraction of each rectangular area as the average value of the pixel gray value minus the standard deviation of the pixel gray value in the area;

step C, selecting the rectangular area with the highest score and further dividing the rectangular area into four rectangular areas;

repeating steps B and C and overlapping the rectangular region with the highest fractionUpdating n times to obtain the final subdivision area Z_b-end；

Obtaining the atmospheric light component image A by using the following formula_c:

|A_c(c∈{r,g,b})|＝min|((Z_b-end-r(x,y),Z_b-end-g(x,y),Z_b-end-b(x,y))-(255,255,255))|，

In the formula (Z)_b-end-r(x,y),Z_b-end-g(x,y),Z_b-end-b(x, y)) is the final subdivided region Z_b-endThe color vector of any pixel point (255 ) is a pure white vector.

Specifically, the deep convolutional neural network is constructed on the base layer image Z_bThe processing to obtain the transmittance image t specifically includes the following steps:

using the first convolution layer to pair the basic image Z_bDown-sampling to obtain low-resolution base layer image Z_b' post-extracting low-level features, wherein the base image Z is aligned using a first convolution layer_bThe formula for downsampling is:

wherein the content of the first and second substances,

the low resolution base layer image Z for the ith convolutional layer at the index of channel c_bThe low-level feature of';

x is a base image Z_bThe abscissa of (a);

y is a base image Z_bThe ordinate of (a);

x' is a low resolution base layer image Z_bThe abscissa of';

y' is the low resolution base layer image Z_bThe ordinate of `;

convolution weight under index channel of c and c' layers for i-th convolution layerAn array;

a deviation vector of the ith convolution layer under the c layer index channel;

σ (·) max (·,0) denotes the ReLU activation function and zero padding is used as a boundary condition for all convolutional layers;

s is the step size of the convolution kernel;

low layer characteristics to be obtained

The number of input layers is n_LExtracting a local feature L from a second convolution layer of 2, wherein the size of a convolution kernel in the second convolution layer is 3 multiplied by 3, and the step size is 1;

low layer feature obtained by laminating a second convolution layer

The number of input layers is n_G1After the third convolution layer with convolution kernel size of 3 × 3 and step length of 2 is input, the number of layers n is input_G2Obtaining a global feature G of the 2 full connection layer;

adding the local feature L and the global feature G, and inputting the sum into an activation function to obtain a low-level feature

Corresponding hybrid feature map F_L＝σ(L+G)；

For mixed feature map F_LObtaining the feature of the lower layer by convolution processing with sigma (L + G)

And (3) performing up-sampling on the corresponding preliminary atmospheric refractive index characteristic graph, and outputting to obtain a transmittance image t (x, y).

Specifically, the global atmospheric light component image A is utilized based on an atmospheric scattering model_cThe transmittance image t and the detail layer image Z_eMisting the originalAnd recovering the image Z, wherein the formula for obtaining the defogged image J is as follows:

wherein J (x, y) is the gray value of the pixel at the point (x, y) in the defogged image J,

Z_e(x, y) is the detail layer image Z_eThe gray value of the pixel at the midpoint (x, y),

Z_b(x, y) is the base layer image Z_bThe gray value of the pixel at the midpoint (x, y),

t (x, y) is the output resulting in a transmittance image,

where t (x, y) is the output resulting transmittance image and η is a constant greater than zero.

In the step, according to an atmosphere scattering model Z (x, y) ═ t (x, y) J (x, y) + A_c(1-t (x, y)), J (x, y) is a defogged image, and (x, y) is the space coordinate of a pixel point, the recovery of the fogging image is carried out, the obtained recovered image can be expressed as,

where t0 is a parameter for ensuring the processing effect of the dense fog region, Z (x, y) ═ J (x, y) + n (x, y), J is a noiseless image, and n is noise, that is, the image processing method is

The noise will be amplified, taking into account that the noise is mainly contained in the detail layer image Z_eAnd noise on the image, the present invention recovers the later image as,

the atmospheric light component a is obtained by step 2.1, the transmittance t is obtained by step 2.2,

in order to reduce the effect of noise on the restored image, which is shown as,

η is a constant, in this example η equals 6, and it is found by experiment that t (x, y) is ∈ [0,1 ∈]When t (x, y) < 1/η, pixel points of the sky region of (x, y) data,

noise of the sky area is prevented from being amplified;

specifically, the pair of mixed feature maps F_LObtaining the feature of the lower layer by convolution processing with sigma (L + G)

And then, performing up-sampling on the atmospheric refractive index characteristic diagram, and outputting a transmittance image t, wherein the method further comprises the following steps:

the loss function is constructed as follows:

L＝L_r+w_cL_c

wherein L is_rTo reconstruct the loss function, L_cAs a function of color loss, w_cTo be assigned to the color loss function L_cWeight of, L_rIs shown as

L_cIs shown as

J is a defogged image, Z is a foggy image, c belongs to (R, G, B) as a channel index, ∠ (J (x, y), Z (x, y)) represents an included angle of three-dimensional color vectors of the foggy image and the defogged image at a pixel point (x, y)), and although a reconstruction error can measure the similarity of the original image and the defogged image, the consistent angles of the color vectors cannot be ensured, so that a color error function is added to ensure the consistent angles of the color vectors of the same pixel point of the image before and after defogging.

The image defogging method and the image defogging device based on the combination of the prior knowledge and the deep learning have the advantages that the weighted guide filter is used for decomposing the foggy image to obtain a detail layer image and a basic layer image, then a quad-tree search method and a deep neural network are used for processing the basic layer image to obtain a global atmosphere light component image and a transmissivity image, and finally the global atmosphere light component image, the transmissivity image and the detail layer image are used for restoring the foggy image to obtain the defogged image. Because the global atmosphere light component image and the transmissivity image are estimated or calculated by utilizing the base layer image, the amplification of image noise is avoided, and the defogging treatment can be better carried out on the foggy image. And calculating a loss function by using the defogged image to perform feedback parameter adjustment, and adding color loss into the loss function to improve the robustness of the algorithm.

In another embodiment of the invention, an image defogging device based on combination of a priori knowledge and deep learning comprises a memory and a processor. The memory is used for storing the computer program. The processor is configured to implement the image defogging method based on the combination of the priori knowledge and the deep learning when the computer program is executed.

The reader should understand that in the description of this specification, reference to the description of the terms "one embodiment," "some embodiments," "an example," "a specific example" or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Those skilled in the art will appreciate that various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims (7)

1. An image defogging method based on combination of priori knowledge and deep learning is characterized by comprising the following steps:

decomposing the input original foggy image Z by using a weighted guide filter to obtain a detail layer image Z_eAnd base layer image Z_b；

Using a quadtree search method to search the base layer image Z_bProcessing to obtain a global atmosphere light component image A_cWherein c belongs to { r, g, b };

constructing a deep convolutional neural network on the base layer image Z_bProcessing is carried out to obtain a transmissivity image t;

utilizing the global atmospheric light component image A based on an atmospheric scattering model_cThe transmittance image t and the detail layer image Z_eAnd restoring the original foggy image Z to obtain a defogged image J.

2. The image defogging method based on the combination of the prior knowledge and the deep learning as claimed in claim 1, wherein the input original foggy image Z is decomposed by using a weighted guide filter to obtain a basic layer image Z_bThe method specifically comprises the following steps:

acquiring a pixel gray value Z (x, y) at any point (x, y) in the original foggy image Z;

obtaining a weighted filter coefficient a at the point (x, y) in the Z of the original foggy image by using the following formula_p(x, y) and b_p(x，y)：

b_p(x，y)＝(1-a_p(x，y))μ_Z,ρ(x,y)；

Wherein, mu_Z,ρ(x, y) is the mean of the pixel grey values of the window centered at point (x, y) in the original hazy image Z, with ρ being the radius;

is the variance of the pixel gray value of the window with point (x, y) as the center and rho as the radius in the original foggy image Z; gamma-shaped_Y(x, Y) is the luminance component of the original hazy image Z at point (x, Y), Y ═ max { Z {_r,Z_g,Z_b}，Z_r,Z_g,Z_bR, G, B values at the point (x, y) in the hazy image Z, respectively; λ is a constant greater than 1;

obtaining the base layer image Z using the following formula_b：

Z_b(x,y)＝a_p(x,y)Z(x,y)+b_p(x,y)；

Wherein Z is_b(x, y) is the base layer image Z_bOf the pixel grey value at point (x, y).

3. The image defogging method based on the combination of the prior knowledge and the deep learning as claimed in claim 1, wherein said base layer image Z is searched by using a quadtree search method_bProcessing to obtain a global atmosphere light component image A_cThe method specifically comprises the following steps:

step A, base layer image Z_bAre equally divided into four rectangular areas Z_b-i(i∈{1,2,3,4})，Z_b-iRespectively has a length and a width of Z_b1/2 for the length and width of;

b, defining the fraction of each rectangular area as the average value of the pixel gray value minus the standard deviation of the pixel gray value in the area;

step C, selecting the rectangular area with the highest score and further dividing the rectangular area into four rectangular areas;

step D, repeating the steps B and C, and carrying out iterative updating on the rectangular area with the highest score for n times to obtain a final subdivision area Z_b-end；

Obtaining the atmospheric light component image A by using the following formula_c:

|A_c(c∈{r,g,b})|＝min|((Z_b-end-r(x,y),Z_b-end-g(x,y),Z_b-end-b(x,y))-(255,255,255))|，

In the formula (Z)_b-end-r(x,y),Z_b-end-g(x,y),Z_b-end-b(x, y)) is the final subdivided region Z_b-endThe color vector at the midpoint (x, y), (255 ) is a pure white vector.

4. The image defogging method based on the combination of the prior knowledge and the deep learning as claimed in claim 1, wherein the construction of the deep convolutional neural network is performed on the basic layer image Z_bThe transmittance image t is obtained by processing, and the method specifically comprises the following steps:

using the first convolution layer to pair the basic image Z_bDown-sampling to obtain low-resolution base layer image Z_b' post-extracting low-level features, wherein the base image Z is aligned using a first convolution layer_bThe formula for downsampling is:

wherein the content of the first and second substances,

the low resolution base layer image Z for the ith of the first convolutional layer at the index of channel c_bThe low-level feature of'; x is a base image Z_bY is the base image Z_bThe ordinate of (a); x' is a low resolution base layer image Z_b'abscissa, y' is the low resolution base layer image Z_bThe ordinate of `;

convolution weight arrays of the ith convolution layer in the first convolution layer under the index channels of the layers c and c';

a deviation vector of the ith convolution layer in the first convolution layer under the c layer index channel is obtained; σ (·) max (·,0) denotes a ReLU activation function and zero padding is used as a boundary condition for all of the first convolutional layers; s is the step length of the convolution kernel of the first convolution layer;

low layer characteristics to be obtained

The number of input layers is n_LExtracting a local feature L from a second convolution layer of 2, wherein the size of a convolution kernel in the second convolution layer is 3 multiplied by 3, and the step size is 1;

low layer feature obtained by laminating a second convolution layer

The number of input layers is n_G1After the third convolution layer with convolution kernel size of 3 × 3 and step length of 2 is input, the number of layers n is input_G2Obtaining a global feature G of the 2 full connection layer;

adding the local feature L and the global feature G, and inputting the sum into an activation function to obtain a low-level feature

Corresponding hybrid feature map F_L＝σ(L+G)；

For mixed feature map F_LObtaining the feature of the lower layer by convolution processing with sigma (L + G)

And (3) performing up-sampling on the corresponding preliminary atmospheric refractive index characteristic graph, and outputting to obtain a transmittance image t.

5. The image defogging method based on the combination of the prior knowledge and the deep learning as claimed in claim 1, wherein the atmosphere scattering model is based on and utilizedThe global atmospheric light component image A_cThe transmittance image t and the detail layer image Z_eAnd recovering the original foggy image Z, wherein the formula for obtaining the defogged image J is as follows:

wherein J (x, y) is the gray value of the pixel at the point (x, y) in the defogged image J, and Z_e(x, y) is the detail layer image Z_eThe gray value of the pixel at the middle point (x, y), Z_b(x, y) is the base layer image Z_bA pixel gray value at a midpoint (x, y), t (x, y) being a transmittance at a midpoint (x, y) of the transmittance image t,

where η is a constant greater than zero.

6. The image defogging method based on the combination of the prior knowledge and the deep learning as claimed in claim 4, wherein the pair of mixed feature maps F_LObtaining the feature of the lower layer by convolution processing with sigma (L + G)

And then, performing up-sampling on the atmospheric refractive index characteristic diagram, and outputting a transmittance image t, wherein the method further comprises the following steps:

the loss function is constructed as follows:

L＝L_r+w_cL_c

wherein L is_rTo reconstruct the loss function, L_cAs a function of color loss, w_cTo be assigned to the color loss function L_cWeight of, L_rIs shown as

L_cIs shown as

J is a defogged image, Z is a fogging image, c belongs to (R, G, B) as a channel index, and ∠ (J (x, y), Z (x, y)) represents the included angle of the three-dimensional color vectors of the fogging image and the defogged image at a pixel point (x, y);

and performing parameter adjustment processing on the deep convolutional neural network by using the loss function.

7. An image defogging device based on combination of priori knowledge and deep learning is characterized by comprising a memory and a processor;

the memory for storing a computer program;

the processor, when executing the computer program, is configured to implement the image defogging method according to any one of claims 1 to 6 based on the combination of the prior knowledge and the deep learning.

Images